Algorithms and Fairness in Lending Markets : When Do Humans and Machines Disagree?

This paper uses a machine-learning-based framework to analyze lending decisions. We develop an adversarial learning model that learns the mortgage lending rules for non-minority borrowers and decompose loan decisions into model explained, favorable disagreement, and unfavorable disagreement segments by comparing lenders’ decisions against model predictions. We apply this approach to HMDA data and illustrate various ways to explore differences across different applicant groups, such as the presence of asymmetric disagreement of loan decisions with ML predictions, linking to ex-ante applicant attributes and to ex-post loan performances. The decomposed loan decisions can further uncover economically meaningful information beyond the aggregated counterpart